Deep Policy: Leveraging Deep Learning for Automated Underwriting and Risk Forecasting in Modern Insurance Models

Abstract

Applications for deep learning within the field of insurance are numerous and varied. A common feature of many of these use cases is the need for deep learning to interact with industry-standard predictive models in areas such as underwriting, price modeling, and risk forecasting. Unfortunately, given the opacity of many deep learning algorithms, actuaries have had limited opportunities to leverage their growing power in these areas, resulting in an artificial divide between the world of traditional models and the emerging world of deep learning. In this paper, we provide illustrative examples of how deep learning can be interfaced with predictive models in insurance to create deep policies, which combine the traditional benefits of policy rules and scoring engines with the power of deep recurrent neural networks for creating accurate, customized models while avoiding issues of opacity and regulatory interpretability.

We then describe a new standard technique of decomposing deep learning models via three simple yet practical steps: (1) Decomposition of Distinctive Risks, (2) Decomposition of Logical Anomalies to provide interpretability for the model's rapid feedback, and (3) Structural Modeling of Sufficient Risk Factors to ensure transparency in model design and regulatory explanation. We then provide examples of the decomposition technique and present structural model descriptions for deep recurrent neural networks and their reflection on visualization colors. Then, we discuss issues such as privacy, data invocation, and the structural coherency of deep recurrent neural networks for creating and using deep policies in the insurance industry before providing examples of detailed abstracts for real insurance applications of the deep policy concept.